AdImpact is a media tracking company that monitors advertisements across thousands of television channels, providing political campaigns and commercial brands with verification that their ads ran as scheduled. Roughly 80% of their business serves political advertisers — a sector where ad placement verification directly influences multi-million dollar campaign decisions.

Their core service depends on one thing: knowing exactly when, where, and what ads aired. For years, that meant a manual monitoring team watching video feeds around the clock. As their channel coverage grew, so did the cost and complexity of that model.

The Challenge 

AdImpact’s manual monitoring operation was a proven approach, but it had an inherent ceiling. Expanding coverage to additional channels required proportional increases in headcount — a cost structure that couldn’t scale alongside the company’s ambitions. They needed a way to automate ad identification and classification without sacrificing the accuracy their clients depended on.

Rhythmic had built the initial AI-powered detection platform for AdImpact earlier in the year — a system that ingests live video streams, identifies commercial breaks, clips the ad content, and delivers structured metadata to AdImpact’s systems. The platform ran through summer on internal data and held up well.

In late October, AdImpact came back ready to go to production. The new requirement: integrating a data partnership with CCR, a media delivery company providing feeds in a different format than the original system was built to handle. What looked like a straightforward production launch turned into a full-scale engineering problem.

The Solution 

Claude was central to this engagement at every stage — from writing and debugging code, to powering the classification engine at the heart of the platform, to helping the team rapidly re-engage with a codebase that had been dormant for months.

On the development side, a significant portion of the platform’s Python codebase was built using Claude — from the initial linear project through the CCR production work. Rather than requiring a team of three or four engineers, Claude’s ability to take well-defined inputs, outputs, and architectural constraints and generate production-quality code allowed Rhythmic to deliver the project with a single engineer. Claude also generated tests to verify individual application functions throughout development.

The production push surfaced three distinct engineering problems that Rhythmic worked through in sequence. First, the infrastructure couldn’t handle the load. When the team initialized the system with live CCR feeds — 40 simultaneous HD video streams plus on-demand content — it saturated the AWS VPC network and overwhelmed the compute instances. Rhythmic scaled up instance sizes and memory capacity to handle the throughput while staying within AdImpact’s infrastructure budget.

Second, CCR’s media delivery format didn’t match the assumptions built into the original system. The team updated the media ingestion layer to accommodate how CCR actually sent content, then re-validated the pipeline end to end.

Third — and most significant — Amazon had updated the AI model powering AWS Bedrock Data Automation during AWS re:Invent, causing ad detection accuracy to drop sharply. Rhythmic diagnosed the issue and turned to Claude. Having tested several models, the team selected Claude via Amazon Bedrock specifically for its superior contextual understanding of the text-based scene metadata the system was analyzing. The pipeline feeds Claude a detailed prompt containing scene descriptions, frame-level video metadata, and audio transcripts cut scene by scene, along with timestamps — and Claude returns a precise determination of where each advertising segment begins and ends.

Claude also proved critical during the troubleshooting phase. When AdImpact flagged quality issues, the team had to re-engage with a codebase that had been untouched for roughly six months. Rather than spending days getting back up to speed, the team used Claude to move through the codebase quickly — making changes and resolving issues in hours.

Because Rhythmic had retained the raw commercial break footage throughout the process, no data was lost during the transition. A reprocessing script fed the saved content back through the updated pipeline, recovering full accuracy across the entire archive.

Technology Used 

• Claude (via Amazon Bedrock)
• Amazon EC2 
• AWS Bedrock Data Automation 
• AWS Lambda 
• Amazon S3 

The Results 

The system reached full production by the end of December. Ad detection accuracy improved from roughly 60% to 98–99%, and at full operation the system was processing 400 to 500 ads per day across live and on-demand content. At last measure, the platform had delivered more than 17,000 discrete pieces of advertising content to AdImpact.

Claude’s role in accelerating development was just as significant as its role in the platform itself. A project that would typically have required a team of three or four engineers was delivered by one — a direct result of Claude’s ability to generate, test, and debug production-quality code throughout the engagement.

The automated pipeline now runs at 20 to 30% lower cost than the equivalent manual monitoring operation — and unlike a manual team, it scales without adding headcount. Rhythmic’s estimate is that the system added $2 to $3 million in enterprise value to AdImpact’s business, before factoring in ongoing operational savings.

Rhythmic continues to manage and optimize the platform in production, handling performance monitoring, updates, and ongoing refinement — the kind of long-term ownership the engagement was designed to support.